Ask a business owner whether their staff use AI tools at work, and the honest answer is usually "probably, I'm not entirely sure." That uncertainty is the definition of shadow AI — the use of AI tools like ChatGPT, Gemini or various free browser extensions to handle work tasks, entirely outside any IT oversight or company policy. It's not staff being reckless. It's staff being efficient, using tools that genuinely help, without anyone having told them where the boundaries are.
The problem isn't AI itself. Used well, it's a genuine productivity gain. The problem is data going into tools nobody vetted, with no visibility into what happens to it afterwards.
How Shadow AI Actually Shows Up
It rarely looks dramatic. It looks like an employee pasting a client contract into ChatGPT to get a faster summary before a meeting. A finance team member pasting a spreadsheet of figures into an AI tool to help draft commentary for a board report. A recruiter pasting CVs into a free AI tool to screen candidates faster. Individually, each of these feels like sensible time-saving. Collectively, across a business with no policy, they represent a steady, invisible flow of potentially sensitive data leaving the organisation's control.
Why "free" matters: Many free-tier AI tools use submitted data to further train their models unless a business explicitly opts out or pays for an enterprise tier with different data handling terms. That means information pasted into a free tool may not just be processed once — it could genuinely become part of the model's training data.
The Specific Risks
- Confidentiality breaches: Client data, contracts, or commercially sensitive information pasted into a third-party tool with unclear data retention and training practices.
- GDPR exposure: Personal data belonging to clients, candidates or staff being processed by a tool the business has no data processing agreement with — a direct compliance gap under UK GDPR.
- Inaccurate outputs treated as fact: AI tools generate plausible-sounding but sometimes wrong answers. Used for a first draft, that's manageable. Used to produce a client-facing figure or a compliance statement without checking, it's a real business risk.
- Loss of institutional control: When AI use is scattered across dozens of individual free accounts rather than a managed business tool, there's no audit trail, no ability to revoke access when someone leaves, and no consistency in how the business actually uses AI.
Why Banning It Doesn't Work
The instinctive response for some businesses is to block AI tools outright at the network level. In practice, this rarely holds — staff switch to personal phones on mobile data, and the business loses visibility entirely rather than gaining control. It also throws away a genuine productivity tool because of a policy gap, which is the wrong trade.
The more effective approach is the same one that eventually worked for shadow IT generally: give staff an approved, well-governed option that's good enough that there's no reason to reach for an unapproved one.
What a Practical AI Policy Looks Like
- Name the approved tools. If your business uses Microsoft 365, Copilot operates within your existing tenant's data boundary and compliance terms — a materially different proposition to a free public tool. Make it the default recommendation and staff have less reason to go elsewhere.
- Define what can and can't go into AI tools, in plain language. Client-identifiable data, financial figures, contracts and anything covered by an NDA should have a clear "don't paste this" line, even into approved tools, unless the tool is specifically vetted for it.
- Require human review of AI output before anything generated by AI goes to a client or gets treated as a final figure.
- Cover AI use in existing security training rather than treating it as a separate, one-off announcement. It should sit alongside phishing awareness and data handling as a standard part of induction and refresher training.
- Review usage periodically. As new tools appear — and they will, constantly — the policy needs to be revisited rather than written once and forgotten.
Where This Sits Alongside Wider AI Adoption
Shadow AI is really a governance problem sitting underneath a genuine opportunity. Businesses that get ahead of it — with a named set of approved tools, clear data rules, and staff who understand both — end up capturing the productivity benefits of AI without the exposure. Businesses that ignore it end up with both: uncontrolled risk and no coherent strategy for using AI well.
This is increasingly part of the conversation we have with clients moving to Microsoft 365 Copilot or similar governed tools — not just "does this save time," but "does this let us finally bring the AI use that's already happening under proper control."
Questions worth asking internally: Do we know which AI tools our staff are currently using? Is there a written policy on what data can and can't be entered into them? Have we given staff an approved, sanctioned option that's genuinely good enough to use instead of a free public tool?
The Bottom Line
Shadow AI isn't a future risk to plan for — for most businesses, it's already happening quietly, every day, in browser tabs nobody's reviewing. The businesses that get this right aren't the ones that ban AI. They're the ones that get ahead of it with a policy, an approved tool, and clear guidance before a problem forces the conversation.